A FRAMEWORK FOR APPLYING

MACHINE LEARNING
TO SYSTEMATIC TRADING
KRIS LONGMORE
ROBOTWEALTH.COM | QUANTIFY PARTNERS
OUTLINE
• Introduction to ML concepts
• ML for trading – pitfalls, where can we go
wrong, why is it so hard – how to not waste
your time!
• One possible framework.
• Some tips to help you make it work.
 INTRODUCTION
• I’m a futures and FX guy, so my examples will reflect that
• Most of this stuff can be applied to other asset classes
INTRODUCTION TO MACHINE
LEARNING CONCEPTS
WHAT IS IT AND HOW IT WORKS, IN A NUTSHELL
INTRODUCTION TO MACHINE LEARNING
CONCEPTS
• What is machine learning? Predictive modelling using statistics.
• Use features/attributes/inputs associated with some event to predict the
outcome of that event.
• We need to uncover the relationship between the features and the target…if
one exists at all.

Learning Predictive
Training Data Algorithm Model
MAKING PREDICTIONS

Out of Sample Predictive

Data Model Forecasts
 The process in a nutshell:

Select New input

target data

Gather Machine
Select Statistical
training learning Forecasts
inputs model
data algorithm

Act on
forecast
Sound simple?
SO WHAT CAN MACHINE LEARNING DO FOR TRADERS?
• Predicting returns is challenging
• Neural networks as universal function approximators – a non-linear
GARCH model? PRICES
• Proxies for price movement from alternative data.
• Improving existing systems – insights into what affects results.
• Train a model to learn your personal preferences for selecting
trading systems (credit: Dr Tom Starke)
• ML for portfolio construction – combine alphas in an optimal manner

• This presentation is going to showcase a framework for forecasting

returns of a trading system
 COMMON PITFALLS
WHAT CAN GO WRONG, AND WHAT CAN WE DO ABOUT IT
TWO BIG PROBLEMS…AND ONE SMALLER ONE
• Tendency to overfit
• Spuriously good results
• Lots of moving parts
Which of these do you think is the smaller problem?
TWO BIG PROBLEMS…AND ONE SMALLER ONE

• Curve fitting bias – we can overcome this with good practice

• Data mining bias – more onerous to account for. Computationally expensive,
not really intuitive. But absolutely critical.
• Lots of moving parts
CURVE FITTING BIAS – SOME PRACTICAL TIPS

• Select parsimonious models

• It may seem counter-intuitive, but reduce the number of features (or use SAE)
• Regularization
• Cross-validate… but consider stationarity and autocorrelation
K-FOLD CROSS VALIDATION

Unfortunately k-fold CV is not a great approach for financial

data which is non-stationary and autocorrelated
TIME SERIES CROSS VALIDATION (TS CV)
• Use a rolling window tested on data adjacent to training data
• Also known as Forward Chaining, Walk Forward Analysis

In practice, the length of the optimization and out of sample

periods may be important factors in a model’s success.

But a robust trading model will be largely insensitive to these.

TIME SERIES CV IS ALSO NOT A PANACEA

• Even if your model passes a TS CV test, there is no guarantee that it won’t just
stop working if and when the underlying modelled relationship changes.
• Particularly true for data-mining or black-box systems.
• At least with model-based systems, you have insight into when it is likely to
break down.
• If you test enough models even with TS CV, you will eventually land on
something that looks really good…but was just lucky
 DATA MINING BIAS (DMB) – WHAT IS IT?

• Essentially, the probability that back-test results are

due to chance.
• Isn’t a good out-of-sample result enough?
• The problem of a finite sample size
• Consider a coin toss. The expectancy of getting heads
is 0.5. But it is not uncommon to have a streak of 2, 3,
4, 5, or more heads in a row. If we only tossed the coin
5 times and got 4 heads, we might conclude that our
coin throws heads 80% of the time.
• If we throw the coin enough times, the sample mean
tends to approach the population mean.
DATA MINING BIAS – WHY IS IT IMPORTANT?

• If we test enough models, we will eventually find something that passes any
out of sample test.
• With machine learning, we can test thousands of models and will certainly
find something that looks good.

• Personally, I think DMB makes ALL the difference in developing machine-

learning based trading systems.
DMB – HOW TO ACCOUNT FOR IT

• There are several approaches

• I will focus on White’s Reality Check (WRC) as it is the one I use most in my
work-flow.
• We will step through the process later.
 TIME FOR SOME FUN:
A MACHINE LEARNING TRADING
MODEL
MACHINE LEARNING DEVELOPMENT FRAMEWORK
Reject or
Pick Pick Pick meta- Construct Test Account
move into
target features features model performance for DMB production

Development process

“Meta-features” may include the choice of learning algorithm, the

number of observations to include in the training data, the timing of
the prediction, the prediction threshold, others.
ANOTHER NUTSHELL

The crux of the process is:

• Which inputs?
• Which outputs?
• How to construct the model? (Which meta-features?)
• Account for DMB
WHAT TO PREDICT?
• My experience is that I get better results when I don’t try to predict returns.

• Better to predict whether a defined positive or negative excursion is hit first –

use a stop order designed to take profit

• Another option is to predict whether an unspecified favourable excursion occurs

– use an initial and trailing stop

• For these examples, I’ll use the initial and trailing stop set up. More aligned with
practical trading.
DIGRESSION…REGRESSION VS CLASSIFICATION

• Regression: predict a numerical value (i.e., result of a trade)

• Classification: predict a binary or multi-class output (e.g., if the market
direction is up or down next week)
• My personal experience is that I get better results when I choose a regression
approach. But many problems can be framed either way and experimentation
is encouraged.
TRADE LOGIC SETUP
• Daily time frame
• Build new long/short models every X days using N previous observations
• Use the model to predict each out-of-sample day’s target – NO IN-SAMPLE PREDICTIONS IN
BACKTESTS!!
• Go long when
LongPrediction > Threshold && LongPrediction > ShortPrediction
• Vice versa for shorts
• Stop and trail mechanisms are volatility-based
• If there is a position open in the same direction then reset the SL as if the position had just
been opened
• Reverse when opposite signal received
• No position sizing - just minimum lots/contracts (not ideal but OK for testing)
SIMPLE EXAMPLE – K-NN APPLIED TO EUR/USD
• For each trading day, predict favourable excursion of long and short trades entered at market
• Model parameters:
• Inputs: last 3 days’ returns
• Learning algorithm: k-Nearest Neighbours with k = 10
• No. observations: 75
• Stop: 50% of ATR(20)
• Trail: 50% of ATR(20)
• We hope to be able to build a relationship where the returns of the last 3 days yields some
predictive power for a trade with our stop/trail mechanism.
 WHAT’S WRONG?
K-NN with k=10
Inputs: 3 days’ returns
Exit: Stop = Trail = 0.5*ATR(20)
• Did we choose the wrong learning
algorithm? (the relationship is too complex
for our model). Different algorithm,
different k?
• Did we choose the wrong inputs? (the inputs
have no relationship with the targets, no
predictive power)
• Are we using the right number of
observations to build our model?
• What about our exit criteria?
EXTENDING THE K-NN EXAMPLE

• Add simple trend filter as input: Close – MA(200)

• Irrelevant inputs only add noise or redundancy – be careful
• Does this additional input improve the model? If not, discard.
K-NN with k=10
Inputs: 3 days’ returns, 1 day’s trend filter
Exit: Stop = Trail = 0.5*ATR(20)
EXTENDING THE K-NN EXAMPLE
• Try using more returns per observation (keeping the trend filter)

5 Returns

10 Returns

20 Returns

15 Returns
 EXTENDING THE K-NN EXAMPLE
• Tune model hyperparameter (k), hold inputs constant
k=3 k=5 k = 10

k = 15 k = 20 k = 25
 EXTENDING THE K-NN EXAMPLE
• How many observations? Vary N holding inputs, k, constant
N = 50 N = 75 N = 100

N = 250
N = 150 N = 200
WHAT HAVE WE DISCOVERED?
• Adding complexity is not always a good thing
• There will often be a relationship between the optimal hyperparameter settings and
the number of observations used
• We have already explored 16 different models – be wary of DMB
• Are there examples of different models that might be accurate under different
market conditions, and therefore can be used in a portfolio?
• What about other methods for combining the predictions of multiple models?
• I don’t have time to delve into ensembles, but you definitely should.
LETS TRY A MORE COMPLEX MODEL

• Support Vector Machine - SVM

• Mostly a classification tool, but can be configured for regression
• The parameter "C" relates to the cost function (it gives weight to the data) and σ to
the mapping of the input space.
• C is like 1/λ if you are familiar with regularization.
• High C  overfit
• Low σ  overfit
LETS TRY A MORE COMPLEX MODEL - SVM
• Train SVM on same data, same inputs
ANOTHER SVM
• Find optimal values of C and σ. Can also try different kernels.
• The values C = 1, σ = 100 produced this equity curve:
 TEST ON OTHER MAJOR CURRENCY PAIRS
GBP/USD USD/JPY

AUD/USD

NZD/USD
EUR/JPY

USD/CAD

Why such different results? What are the implications?

THE EFFECT OF SAMPLING TIME

• So far our models are built using daily closing prices.

• FX is a 24-hour market, so the choice of closing time is somewhat arbitrary.
• Train same model using same inputs but construct daily time series at different
GMT offsets. Be cognizant of daylight savings time.
• Do results randomly fluctuate with Greenwich Mean Time offset, or do they
deteriorate systematically as you move away from the best offset?
• This effect is also present in futures markets.
THE EFFECT OF SAMPLING TIME
USD/JPY @ 0100 GMT AUD/USD @ 2000 GMT

USD/CAD @ 1000 GMT

GBP/USD @ 1700 GMT

FILTER ON MAGNITUDE OF PREDICTION

• Only enter trades when prediction is greater than some threshold

• Rather than remain constant, threshold should vary with underlying volatility
• If performance degrades, we unlikely have an edge in terms of predicting the
magnitude of the next trade…but we may have an edge in predicting the
directional accuracy and the distribution of these predictions.
 FILTER ON MAGNITUDE OF PREDICTION
Previous GBP/USD SVM model,
but filter small predictions.
PORTFOLIO OF SVM MODELS

We could certainly create better

portfolios by selecting better
features/meta-parameters for each
market.
 IDEAS FOR MODEL CONSTRUCTION

Inputs: Targets: Learning Algorithms:

• Returns of different assets • Fixed profit target • k-NN

• Favourable excursions • SVM/SVR
• Volatilities of different assets
• Bullish/bearish (classification) • Boosting machines
• Time of day
• Returns • Decision trees
• Trend indicators • One step ahead • Random forest
• Fundamental data • Multiple step ahead • Neural networks
• Alternative data • Intra-day returns • Deep learning techniques
BEWARE THE TEMPTATION TO TEST EVERYTHING
Between the choice of target, features and meta-features, there are almost unlimited possible
models to choose from. Hence why DMB can be so insidious with this approach.

If you try to test everything, you will thwart yourself with data mining bias and the limits of
practical computational power.
WHY DMB CREEPS IN

How many possible models? Say we have:

• 4 possible inputs, of which we choose at least 2
• 4 possible values each for C and σ
• 3 options for the target denoted by different stop/trail mechanisms
• 4 options for the number of observations – say 50, 100, 150, 200
That’s over 2,000 possible models, and we haven’t even considered other learning
algorithms yet.
OTHER LEARNING MODELS
Most important thing is to be aware of the limitations and requirements of your learning
algorithm – they are all different!
Example: Random Forest
• Good at not being influenced by non-important variables. Can be used for feature selection.
• Non-deterministic – can give different results for each simulation due to random initialization
• Doesn’t do so well when we use few inputs - tends to overfit in this case.
Example: Neural Network
• Easy to overfit with complex networks
• Features should be scaled and normalized
• Non-deterministic due to random initialization
ADAPTING MODEL PARAMETERS
• The model ideas presented so far seem quite arbitrary in their choice of parameters
• The parameters are also static
• What if we adapted these parameters throughout the life of the strategy?
• What does that set up look like?
ADAPTING MODEL PARAMETERS
• At regular intervals, say weekly, build N different models and simulate their
performance over the previous 100 (say) days. Select the best model for
trading in the next period.
• Computationally expensive
• Test the performance of N models during the past 100 days
• Each individual model is also retrained (say, weekly) during those 100 days
ADAPTING MODEL PARAMETERS
Dynamic model across the input and meta-
TRAINING TRADING parameter space as well as model training.

Model 1 Inputs, targets number of observations, etc.

change dynamically as the market evolves.
Model 2
A nice by-product of this approach is that it
Model 3 gives insight into how the market is evolving
with time.
Model 4 Best model
You also get feedback on which markets are
more predictable at any given time. You will
… see that sometimes very few or even zero
models show acceptable predictive utility.
Model N
Other times you will get many models to
choose from – you can even select the best
performers to put in an ensemble.
ACCOUNTING FOR DATA MINING BIAS

• How do we distinguish profitable results from the results our machine learning
model can give due to random chance? (how to measure data mining bias?)
WHITE’S REALITY CHECK
• White, H. 2000. A Reality Check for Data Snooping, Econometrica, Vol.68, No.5
• A good description in Aronson, 2006. Evidence Based Technical Analysis
• Further improvements proposed by Hansen (2005), Romano and Wolf (2005), and Corradi
and Swanson (2011).
• Available online for free (except Aronson).
• Reference list available at http://robotwealth.com/recommended-reading/
• We’ll step through the actual process using the back tests I ran in creating this presentation.
WHITE’S REALITY CHECK (WRC) – WHAT IS IT?

Definition:
A process for creating the distribution of the best variants of the systems tested
during the development process, under the null hypothesis that all tested systems
have an expectation of zero.

• What does that actually mean?

WRC – WHAT IS IT?

• WRC allows us to see where our model’s performance lies in relation to the
randomly good, bad or indifferent results that could have been created
through our particular development process, if they had zero expectancy.
• We want our model to beat the vast majority of these systems.
EXAMPLE OUTPUT

• WRC results in an empirical distribution of “best”

strategies from our development process under H0
that all variants tested have an expectation of
zero.
• In this case, our strategy beats roughly 98% of
“best” strategies.
• Roughly 2% chance of getting this result if H0 is
true.
• Depending on our desired confidence level, we
may reject H0 and conclude it is unlikely our
strategy is merely lucky.
CONFUSED?

• If this sounds confusing, showing you the process might clarify things.
DIGRESSION - BOOTSTRAP
• The empirical distribution is constructed via bootstrapping (with replacement)
the returns series of each variant tested in the development process.
• Bootstrapping is sampling with replacement. This means that one observation
can appear more than once in the bootstrapped sample.

Original 0.24 2.23 0.19 -0.09 0.45 -0.58

Bootstrap 2.23 2.23 2.23 0.45 -0.09 -0.09

Randomized 0.19 -0.09 0.24 2.23 -0.58 0.45

WRC – STEP 1

Develop your strategy, storing the returns series of all variants tested. Pick the
best performer and note its performance (profit factor, sharpe, etc).
Of all the strategies tested on EUR/USD, a SVR model returned a Profit Factor
of 1.51
WRC – STEP 2
Create a zero-expectancy returns series from each of the stored returns series.
Simply subtract the mean return of each variant from each day’s return.
WRC – STEP 2
Returns Returns – mean(Returns)

This results in a series of balance curves with the same characteristics as

the tested variants, but with zero expectancy.
WRC – STEP 3
Resample each zero-mean curve using bootstrap.
This results in a new set of balance curves, many with non-zero final balances.

Returns – mean(Returns) Bootstrapped Zero-Mean Returns

WRC – STEP 4

Select the best performer from these resampled curves and note down its
performance.
Bootstrapped Zero-Mean Returns
Best Profit Factor = 1.35
 WRC – STEP 5

Repeat steps 3 and 4 a few thousand times.

Bootstrap 1: Bootstrap 2: Bootstrap 3: Bootstrap 4: Bootstrap 5000:

Best PF = 1.7 Best PF = 1.2 Best PF = 1.3 Best PF = 1.2 Best PF = 1.5
WRC – STEP 6
Construct a frequency histogram from the several thousand values for the best
performer in each iteration of the bootstrap exercise.

Median ~ 1.2
Long right tail out to ~ 1.7
WRC – STEP 7

Work out where your selected strategy’s performance sits on that distribution.
The percentage of bootstrapped values greater than your strategy’s value is
the p-value for the statistical test for H0. If 1-p > confidence level, reject H0.

Benchmark PF = 1.51, which beats 98% of bootstrapped performance curves

p = 0.02, so we reject H0
WRC – SOME PRACTICALITIES

• Set up step 1 at the outset – saving all the returns series throughout the
development process.
• Easy solution: include a function in your back-test code that appends the daily
returns series of any back test as an array to a file with the same name as the
strategy.
WRC – SOME PRACTICALITIES

• WRC tends to Type II errors.

• That is, tends to reject good strategies, particularly when many variants are
tested.
• Intuitively that makes sense – the more variants we test, the more likely we are
to get lucky.
• An improvement proposed by Romano and Wolf (2005) may help.
WRC – SOME PRACTICALITIES

• When variants have a negative expectancy, WRC tends to further over-

estimate data mining bias.
• An improvement proposed by Hanson (2005) may help.
FULL DISCLOSURE

Looks like a really nice result, but in reality I kind

of knew where to look because of the large body
of research I’ve previously undertaken in this area.

Can we therefore have confidence in this data

mining bias accounting procedure?
ONE OF MANY ML DEVELOPMENT FRAMEWORKS
Reject or
Pick Pick Pick meta- Construct Test Account
move into
target features features model performance for DMB production

Development process

Welcome to the land of opportunity!

Automate the process

Highly accessible

You can spend a lot of time on this, please don’t waste

that time. Account for data mining bias!
IB Account Offerings
• Lowest Margin Rates and Commissions*
• Security Financing - real-time depth of availability and rates help protect against buy-ins and recalls.
• Safety of Assets strong balance sheet, large relative equity capital, client fund protection
• Innovative trading technology; trade in over 100 market centers in 24 countries, direct market access to stocks,
options, futures, forex, bonds, ETFs and CFDs from a single account.
• Speed of Execution and
• Risk Management.

IB API
Build a complete trading application that connects to our advanced order routing and trading system using our IB API:
• Choose from among several available IB API programming languages, including Java, .NET (C#), C++, ActiveX or DDE.
• FIX CTCI – For clients with the knowledge of and resources to support FIX CTCI, a robust, industry-standard solution.

Interactive Brokers LLC is a member NYSE, FINRA and SIPC. Supporting documentation for any claims and statistical information will be provided
upon request. *According to the Barron’s 2016 online broker review. For additional information see, ibkr.com/info